Rankings were a fair proxy for visibility while every search ended with a click on a link. That is no longer how many buying questions end. They end in a chat window, with a short answer and three or four sources behind it, or no source named at all.
Creeper’s LLM Citation Benchmark puts a checkable number on what happens in that window. This page explains how the number is produced and what it does not prove.
Why rankings do not carry over into answers
A results page offers a list of options, and the click goes to whoever sits highest. A model reads the web, picks a few passages, and writes one reply. Being first on Google is strong evidence that Google can crawl, index and trust a page. It is weak evidence that a language model will name that page when a buyer asks a question in your category.
Keyword volume measures how often people search a phrase. It says nothing about whether a model treats your page as a source. Backlinks still work as corroboration and still matter for classical ranking, but a strong link profile does not predict citation. Well-linked pages go unquoted all the time, and thinner pages that state facts cleanly get picked up instead.
AI answers add a second problem: they are not stable. The same prompt can return different sources an hour apart, and different models disagree with each other. A claim like “we are visible in ChatGPT” is worth very little until someone defines what was asked, to which model, how many times, and what counted as visible.
What counts as a citation
A citation is a source the model names or links in its answer to one of your benchmark prompts. We count two kinds and never merge them:
- Source citation. Your domain appears in the list of sources attached to the answer.
- Brand mention. Your company or product is named in the body of the answer, with or without a link.
We resolve redirects, treat copies of a syndicated article as one source, and discard links to pages that only quote your press release. Five figures come out of each run:
| Metric | The question it answers |
|---|---|
| Citation rate | On what share of prompts does the model name you at all? |
| Voice share | Of all sources cited across your prompt set, what share are yours? |
| First-source rate | How often are you the first source the answer leans on? |
| Accuracy | Does the model describe your product, price and coverage correctly? |
| Source mix | Which domains keep appearing, and who owns them? |
How a benchmark run is built
1. The prompt set comes from real questions. We collect queries your pages already receive, questions your sales team answers on calls, and the category questions buyers type into assistants in their own words. Prompts are grouped by type: brand, category, comparison, problem. A single strange answer then moves one group, not the whole score. The set is fixed and versioned. Changing it resets the baseline, and every version change is dated in the report.
2. The same prompts go to a fixed panel of models. Typically ChatGPT, Perplexity, Gemini, Copilot and Claude, chosen by which assistants your market actually uses. Wording stays identical across models, the panel runs inside one week, and no account personalization is allowed into the results.
3. Each prompt runs more than once. Models sample their output, so one answer is an anecdote. Every prompt is repeated per model per cycle and we report the spread, including prompts where the model cited you once out of five. A benchmark that reported the best run would be a screenshot generator.
4. Citations are matched, then checked. An automated pass matches your domain and brand variants, then a person reviews a sample of every cycle’s matches. This catches the two errors that flatter a report: counting your own press release as independent corroboration, and counting a competitor whose name resembles yours.
5. Scores aggregate, transcripts stay attached. Results roll up by prompt group, by model and by cycle, placed next to a competitor set you name. Every score links back to the answer that produced it. If you want to argue with a number, you can read the exchange and argue with the model instead.
What the number cannot tell you
It is not a traffic forecast. Being cited and being clicked are different events. An answer engine can send readers without a citation, and cite you without sending anyone.
It does not establish causation on its own. Models are retrained, retrieval indexes change, and a jump in citation rate the week after a content edit may have nothing to do with that edit. Repeated over several cycles, a benchmark makes changes legible. It does not isolate a single cause.
The mechanism is opaque. Nobody outside the model labs knows which passages get retrieved or why. We measure outcomes, not an internal ranking, because no such ranking is readable from outside.
Silence is ambiguous. A model can use your content without naming it. A missing citation means you were not credited, not that you were not read.
What the benchmark changes in practice
When the number is low, the work is usually unglamorous. State the answer to the buyer’s question in a sentence a machine can lift. Describe the company the same way on every page, in the same words. Keep product, price and coverage facts current and in structured data. Leave AI crawlers welcome in robots.txt. Get the claim corroborated somewhere you do not own, because a fact that appears on one domain is easier to skip than one that three independent sites repeat.
Then measure again on the same prompt set. Run, fix, rerun, compare.
Where this sits next to your technical audit
The free SEO audit answers whether search engines and AI crawlers can reach, parse and render your pages. Our real user monitoring answers whether those pages hold up for people once they arrive. The benchmark answers the layer above: given access, does a model choose your material and credit it. A citation problem is often a crawl or markup problem wearing a different hat, which is why we run the audit alongside the benchmark.
Start with a baseline
A baseline run gives you the prompt set, the model panel, the transcripts and your citation rate before anyone edits a page. If the number is already healthy, we will say so and tell you where the money is better spent.